Face detailer for animatediff resulting in black mask.

[amnesicloud]

Hi there, thanks for the tutorials, it has been a great help.
I'm having problem when making modifications based on your workflow: ComfyUI-Impact-Pack/workflow/detailer-for-animatediff.png
https://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/workflow/detailer-for-animatediff.png

I'm using this gif as input:

and this is my workflow, where I removed the first Ksampler node since I'm using specific input:
test_workflow (1).json

and my result are as follow:

where the face is black masked. I'm wondering what's causing this problem?

[amnesicloud]

to me it has successfully detected the face, but failed to redraw, but I can't make it work

[ltdrdata]

What is your gpu?

[amnesicloud]

using Apple Silicon M2, I use comfyui with this command:
./venv/bin/python main.py --force-fp16

[amnesicloud]

with 32G Mem

[ltdrdata]

Show me the full log comfyui.log

My guess is that M2 is related to issues occurring when performing ksampling at a specific resolution.

[amnesicloud]

update: I adopted the method from Kosinkadink/ComfyUI-AnimateDiff-Evolved#48 (comment), and the black mask problem is gone.
However now the face part is weird images, it seems

the log is as follow
comfyui.log

[ltdrdata]

Before proceeding further, it is recommended that you update your ComfyUI. It's already a month old version.

[amnesicloud]

I have updated ComfyUI, and I get rid of previous update, it causes problem in animatediff workflow
comfyui.log

current result is like this:

[ltdrdata]

Try creating images in 640, 674 resolution in the basic AnimateDiff workflow without FaceDetailer, if that's not an issue, apply i2i.

[amnesicloud]

I'm not sure what you mean i2i, image2image?
I have "basic AnimateDiff workflow without FaceDetailer" working, and I've tried to apply face detailer on each frame (is that what you meant by "i2i"?), however that would cause the face flickering, that's why I'm looking for an animatediff detailer solution.

[ltdrdata]

I meant vid2vid. Just make source image size as 640, 674.